Prashan Madumal is a PhD candidate in the University of Melbourne associated with the AI and Autonomy Lab at the School of Computing. His research interests include explainability of agents, reinforcement learning, causal learning and human-agent interaction.
Explainable Artificial Intelligence, a research agenda investigated since the era of expert systems saw a renewed interest in recent years. But much research and practise give little attention to the human as the systems’ end-user and are done in a vacuum, often ignoring the vast body of literature on explanation and explainability in philosophy, cognitive science and psychology. By unifying theories of causality in cognitive science and AI, I introduce a new explainable model for reinforcement learning agents. A causal model can itself be interpretable, and to provide explanations, I propose new methods for how explanations can be generated, selected and contrasted. I also briefly discuss how causal models can be extended to include different types of causal explanations.